AI Layoffs Are Changing the Economics of Pregnancy in China — and America Has the Opposite Problem
One of the stranger responses to AI-era job insecurity is a pregnancy test.
Chinese reporting has documented women bringing pregnancies forward when redundancies loom, a tactic known as zhànlüè huáiyùn, or “strategic pregnancy”. The least sensational detail is probably the most useful. In interviews published by The Paper, a lawyer who had advised women considering the tactic said most were not deciding to have children solely because of a threatened layoff. They already wanted a child, often a second one, and moved the timing forward because their employment prospects had changed. (The Paper)
That makes strategic pregnancy less a story about women gaming maternity law than a story about timing. A protection that is valuable in ordinary circumstances becomes much more valuable when a redundancy is expected within months.
The same employment shock produces a different calculation in the United States. Chinese law can make some forms of dismissal materially harder during pregnancy. American law protects workers from pregnancy discrimination but generally does not remove them from an otherwise lawful reduction in force. As companies rely more heavily on automated measures of output, a further problem appears: lawful absence can leave behind data that looks like poor performance.
Technology may create the employment shock. The law around the worker decides what options she has when it arrives.
The option value of pregnancy
China's protection is substantial, although it is narrower than some accounts of strategic pregnancy suggest.
Article 42 of the Labor Contract Law prevents employers from using the dismissal routes in Articles 40 and 41 against employees during pregnancy, the post-birth period and breastfeeding. Those routes include dismissal for continuing incompetence after training and certain large-scale economic redundancies. A fixed-term contract that would otherwise expire is also extended while the Article 42 condition continues. (Supreme People's Court of China)
Article 39 still permits dismissal on specified grounds, including serious violations of company rules, serious dereliction of duty and, during probation, being proved unqualified for recruitment. Pregnancy therefore blocks important routes to redundancy; it does not prevent every lawful termination. (Supreme People's Court of China)
The compensation arithmetic also needs care. Article 87 provides twice the Article 47 compensation rate for unlawful termination. Article 47, however, caps the salary base for high earners at three times the relevant regional average wage and caps the applicable service period at twelve years. The popular claim that a senior technology worker necessarily receives 2N calculated on uncapped actual salary does not follow from the statute. (Supreme People's Court of China)
The incentive remains. Suppose a woman already expects to have a child within two years and believes her employer will restructure in six months. Bringing the pregnancy forward can move her from an ordinary redundancy pool into a protected category just when that protection has its highest value.
The key decision is often timing, not whether to become a parent at all.
That is also why the most dramatic version of the story, in which layoffs somehow solve China's fertility crisis, is hard to sustain. China recorded 7.92 million births in 2025 and its population fell by 3.39 million. The government has introduced a nationwide subsidy of RMB3,600 a year for each child under three. Whatever strategic pregnancy does at the margin, it has not reversed the country's demographic decline. (National Bureau of Statistics of China)
A narrower claim fits the evidence better: fear of unemployment may alter the timing of births that were already under consideration. Labour law can make that timing financially consequential.
America prices the same risk differently
The American version of the trade is much weaker.
Federal law prohibits pregnancy discrimination and requires qualifying pregnancy-related accommodations. The Equal Employment Opportunity Commission includes layoffs among the employment decisions in which pregnancy discrimination is unlawful. (EEOC)
But the protection is against discriminatory treatment. It does not ordinarily stop an employer from including a pregnant worker in a reduction in force when she would have been selected regardless of the pregnancy.
Eligible workers can receive up to 12 weeks of job-protected FMLA leave, generally unpaid. The Department of Labor says an employee on FMLA leave has no greater right to the job than she would have had without taking leave. If an employer can show that she would have been laid off anyway, the right to reinstatement does not save the position. The United States also still has no federal guarantee of paid family and medical leave for private-sector workers. (U.S. Department of Labor)
Pregnancy discrimination is nevertheless common enough to matter. A 2022 national survey commissioned by the Bipartisan Policy Center and Morning Consult found that 20% of mothers reported experiencing pregnancy discrimination at work. (Bipartisan Policy Center)
What the United States lacks is credible evidence of a widespread practice of using pregnancy as redundancy insurance. The incentives do not support it. Nor does having a child obviously make a woman a safer hire. Research using 811,000 administrative earnings histories found large and persistent motherhood penalties, including at firms with female leadership and among couples in which the woman had been the higher earner before childbirth. (PNAS/PMC)
Economic insecurity pushes fertility in the same direction. A 2026 Demography study using a cohort-discontinuity design found negative causal effects of the Great Recession on fertility across generations of U.S. women, with larger effects among younger cohorts. (PubMed)
In the United States, fear about work tends to delay childbearing rather than turn pregnancy into an employment hedge. The fear can be similar. The institutional response is not.
The denominator problem
AI adds a different risk to the American model.
In July 2026, 26 current and former Meta employees brought claims alleging that AI-assisted workforce systems had contributed to a layoff process that disadvantaged employees who had taken protected medical or family leave or received disability accommodations. Meta denied that AI made the termination decisions. The company told the court that human business leaders selected employees using documented criteria and that no AI-assisted performance scoring was used for the reduction in force. (U.S. District Court order)
The judge denied the employees' request for a temporary restraining order, so the ruling was not a finding that Meta had discriminated. He did conclude that the record raised “serious questions” about the merits that required further evidence. Employee declarations cited in the order alleged that AI-usage metrics fell during periods of leave and that AI tools were being integrated into future performance-review processes. (U.S. District Court order)
On September 2, WIRED reported that Meta had told employees their performance evaluations would no longer depend on how much they used AI tools, while the company continued to encourage broad AI adoption. The policy change is not an admission about the lawsuit or the earlier layoff process. It does show how quickly a measurement choice can become an employment-law question. (WIRED)
The denominator is the problem. Take a quarterly productivity measure such as code committed, tickets closed, customer conversations completed, messages answered or AI tokens consumed. An employee who spends eight legally protected weeks away from work will produce less during that quarter. A crude ranking system records lower output. A properly designed one adjusts for the time the employee was actually available to work.
No explicit pregnancy field is necessary for the bad outcome. A system that treats protected absence as zero productivity rather than missing exposure can turn lawful leave into a negative performance signal.
That risk is broader than model bias in the usual sense. It sits in the design of the measurement itself: the period being scored, the activity the system can observe, the treatment of absence, and the kinds of work that leave a machine-readable trail. An algorithm can be neutral about sex and pregnancy and still produce a distorted ranking if the denominator is wrong.
A legal right to be absent loses much of its force if the performance system later treats the absence as evidence that the worker produced less.
Labour law belongs in the system design
Companies that use AI in performance management, workforce planning and restructuring will have to treat employment law as part of the system design, not as a review conducted after the model has been built.
That means adjusting denominators for protected leave, keeping auditable decision trails, testing how protected absence affects rankings, and preserving meaningful human review. Vendors that can show that their systems distinguish poor performance from periods when an employee was legally entitled not to work should be easier to trust than systems built around raw activity counts.
The investment question follows from the same point. A company can report impressive AI-driven productivity while its internal metrics create litigation risk, alienate employees or push managers toward bad talent decisions. Headcount efficiency is not the same thing as operating efficiency if the measurement system cannot tell the difference between weak performance and protected absence.
China shows the opposite distortion. Strong pregnancy-linked dismissal protections can transfer substantial employment risk from the worker to the individual employer. Rational employers may try to price that expected cost before pregnancy occurs through hiring preferences, promotion decisions or suspicion of women they expect to take maternity leave. Such discrimination can itself be illegal and harmful. The incentive still exists.
That matters for policy design. If a government wants more children but leaves a large share of the cost of childbirth with whichever employer happens to employ the mother, firms have a reason to avoid that cost. Spreading more of it through insurance, taxation and public provision reduces the incentive for an individual employer to treat potential motherhood as a balance-sheet liability.
China's national childcare subsidy moves some cost in that direction. Its demographic numbers show how much further the problem extends. Governments can pay parents cash to have children, while employers continue to price the same decision through job security, promotion, earnings and later employability.
The Chinese and American cases are useful together because they expose two different failures. In China, strong protection can make the timing of pregnancy economically valuable during a downturn and can also sharpen employers' incentives to discriminate before pregnancy occurs. In the United States, narrower protection leaves lawful layoffs available while automated productivity systems can misread protected absence after it occurs.
These are rational responses to rules, metrics and costs. Blaming women for exploiting maternity protections, or blaming algorithms as though they possess motives, misses the mechanism.
The harder question for the AI workplace is whether its systems can distinguish low output from a legally protected interruption in output. Pregnancy is only the clearest test. Illness, disability, caregiving and recovery create the same problem.
If the system cannot make that distinction, employment law will exist on paper while the data layer quietly prices people as though it does not.